Tilt-Stable Local Minimizers

• Tilt-stable local minimizers are local solutions that remain uniquely optimal and Lipschitz stable under small linear perturbations of the objective.
• They are characterized by uniform quadratic growth, strong metric regularity, and positive-definite second-order conditions that link sensitivity to robust performance.
• Their theory extends across smooth, nonsmooth, composite, and conic optimization, underpinning effective generalized Newton methods and convergence analysis.

Tilt-stable local minimizers are local solutions whose behavior under small linear perturbations of the objective is both unique and Lipschitz controlled. In the standard variational-analytic formulation, one studies the localized argminimum mapping for perturbed problems of the form $f(x)-\langle v,x\rangle$; tilt stability holds when this mapping is single-valued near the unperturbed problem and varies Lipschitz continuously with the perturbation. Across the literature, this notion is linked to uniform quadratic growth, strong metric regularity of the subdifferential, and positive-definiteness of generalized second-order objects, and it has been extended from smooth finite-dimensional models to prox-regular nonsmooth functions, Banach-space growth theory, weakly qualified nonlinear programs, nonpolyhedral conic problems, composite models, and spectral matrix optimization (Mordukhovich et al., 2011, Drusvyatskiy et al., 2013, Chieu et al., 2017, Mordukhovich et al., 15 Jul 2025).

1. Perturbed minimization and the classical notion

The classical definition uses an extended-real-valued objective $f$ and a reference point $\bar x$. For some $\gamma>0$, the localized perturbation map is

$$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$

A point $\bar x$ is a tilt-stable local minimizer if $M_\gamma$ is single-valued and Lipschitz continuous around $v=0$, with $M_\gamma(0)=\bar x$. This is the formulation used in the finite-dimensional nonsmooth literature, in generalized Newton methods, and in conic optimization (Chieu et al., 2017, Mordukhovich et al., 2020, Benko et al., 2018).

The definition is quantitative. One says that $\bar x$ is tilt-stable with modulus $f$0 if the localization is Lipschitz with constant $f$1. The exact tilt bound is then defined by

$f$2

or by equivalent infima over localized Lipschitz moduli of the argminimum map (Chieu et al., 2017, Chieu et al., 9 Aug 2025).

This perturbational viewpoint is stronger than mere isolated local minimality. It encodes not only that $f$3 minimizes the unperturbed objective locally, but also that nearby tilted problems admit a uniquely selected nearby minimizer. In that sense, tilt stability is a local sensitivity property rather than only a first- or second-order necessary optimality condition (Mordukhovich et al., 2011, Mordukhovich et al., 2020).

2. Growth, subdifferentials, and second-order characterizations

A central theme in the theory is that tilt stability is equivalent to quantitative growth and regularity properties. For prox-regular and subdifferentially continuous functions in finite dimensions, tilt stability is equivalent to strong metric regularity of the limiting subdifferential around $f$4, and also equivalent to positive-definiteness of the second-order subdifferential: $f$5 The same finite-dimensional theory shows equivalence with metric regularity plus positive-semidefiniteness and trivial kernel of the generalized Hessian (Drusvyatskiy et al., 2013).

This second-order viewpoint becomes explicit for amenable composites. For objectives of the form

$f$6

second-order subdifferential chain rules compute $f$7 from the smooth data and the second-order geometry of the outer function. In smooth nonlinear programming under LICQ, the resulting criterion reduces to the strong second-order optimality condition for the Lagrangian, so tilt stability becomes equivalent to SSOC (Mordukhovich et al., 2011).

A complementary characterization uses the subgradient graphical derivative. Under prox-regularity and subdifferential continuity, tilt stability is equivalent to a neighborhood-uniform positive-definiteness condition: $f$8 for nearby $f$9. This formulation replaces a single pointwise Hessian test by a neighborhood condition on tangent directions to the graph of the subdifferential (Chieu et al., 2017).

Recent Banach-space work shows that this circle of ideas persists beyond quadratic growth. For a proper lower semicontinuous function on a reflexive real Banach space and a strict local minimizer $\bar x$0, local $\bar x$1-growth,

$\bar x$2

is equivalent to a monotonicity-type condition for minimizers of tilted problems, to a Hölder-type “tilt sub-stability” estimate

$\bar x$3

and to a Łojasiewicz-type inequality with linear perturbations

$\bar x$4

where $\bar x$5 (Corella et al., 2024). The same paper presents a global equivalence between growth and a subdifferential error bound involving $\bar x$6, and treats this as an analog of the Polyak–Łojasiewicz condition with the gradient replaced by a linear tilt (Corella et al., 2024).

In the semi-algebraic setting, isolated local minimizers admit a tangency exponent $\bar x$7 such that $\bar x$8-th order sharp local minimality, $\bar x$9-th order strong metric subregularity of $\gamma>0$0, and a Łojasiewicz gradient inequality with exponent $\gamma>0$1 are equivalent whenever $\gamma>0$2 (Pham, 2019). This sits close to tilt-stability theory because it organizes the same growth–subdifferential–Łojasiewicz triad on a higher-order scale.

A recurring nuance is that these second-order positivity conditions characterize tilt stability, not arbitrary local minimality. In particular, positive-semidefiniteness of the generalized Hessian is not necessary for local optimality in full generality (Drusvyatskiy et al., 2013).

3. Nonlinear moduli and local manifold structure

The classical Lipschitz/quadratic framework has been generalized by replacing linear and quadratic gauges with admissible functions. In this setting, a proper lsc function on a Banach space has a $\gamma>0$3-tilt-stable local minimum at $\gamma>0$4 if there exist $\gamma>0$5 and a mapping

$\gamma>0$6

such that $\gamma>0$7 is a local minimizer of the tilted problem and satisfies the nonlinear stability estimate

$\gamma>0$8

This is paired with $\gamma>0$9-stable local well-posedness, a growth condition for tilted objectives, and the two notions are equivalent when

$$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$0

For $$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$1 and $$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$2, the theory recovers the standard quadratic-growth/tilt-stability equivalence (Zheng et al., 2016).

The same admissible-function framework relates stability to subdifferential regularity. Strong metric $$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$3-regularity of $$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$4 is sufficient for $$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$5-stable local well-posedness, while a localized convexified subdifferential regularity condition is necessary. In the convex case, $$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$6-stable local well-posedness is equivalent to strong metric $$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$7-regularity of $$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$8 (Zheng et al., 2016).

Tilt stability also has geometric consequences for nonsmooth structure. Under prox-regularity, quadratic minorization, and a tilt-stable local minimum, a $$M_\gamma(v):=\operatorname{argmin}\{\, f(x)-\langle v,x\rangle \mid x\in \mathbb{B}_\gamma(\bar x)\,\}.$$9-type decomposition separates directions of nonsmoothness from directions of smoothness. The second-order component

$\bar x$0

acts as a smooth tangent candidate, and a localized minimizer selection $\bar x$1 defines a manifold

$\bar x$2

On this manifold, the convexified function and the original function coincide locally, and the restriction becomes $\bar x$3; under the “fast track” condition $\bar x$4 and additional assumptions, $\bar x$5 is continuously differentiable and $\bar x$6 is a $\bar x$7-smooth manifold (Eberhard et al., 2016). This places tilt stability near partial smoothness, manifold identification, and smooth reduction of nonsmooth problems.

4. Constrained, conic, composite, and matrix optimization

For nonlinear programming, the modern theory replaces classical LICQ-based results by much weaker qualification regimes. One line of work shows that under MSCQ, tilt stability follows from pointwise second-order conditions expressed through the Lagrangian Hessian and appropriate multiplier sets, and in particular that SSOSC guarantees tilt stability at stationary points under MSCQ (Chieu et al., 2017). A sharper point-based theory uses the pair MSCQ + BEPP, derives explicit second-order formulas for the indicator of the feasible set, and obtains complete characterizations of tilt-stable minimizers in terms of extreme multipliers in critical directions, together with exact tilt bound formulas (Gfrerer et al., 2015). More recently, under relaxed constant rank constraint qualification, point-based characterizations and an explicit exact bound were derived without requiring linear independence of equality-constraint gradients (Chieu et al., 9 Aug 2025).

For second-order cone programming, complete neighborhood and point-based characterizations have been established under MSCQ. The second-order test involves the Hessian of the objective, the Hessian of the constraint mapping weighted by multipliers, and a curvature term $\bar x$8 that reflects the nonpolyhedral geometry of the Lorentz cone. The analysis splits into out-of-kernel and in-kernel regimes, and the in-kernel case brings in 2-regularity as an additional structural condition (Benko et al., 2018).

For nonlinear semidefinite programs with convex feasible sets, tilt stability has been analyzed via the second subderivative of the extended-valued objective

$\bar x$9

Point-based sufficient characterizations are available without constraint nondegeneracy by using multiplier restrictions and a second-order formula for $M_\gamma$0; in the linear positive semidefinite cone constraint case, one also gets a necessary characterization, and under a suitable restriction on the multiplier set, a sufficient-and-necessary point-based criterion (Liu et al., 2024).

For general composite problems

$M_\gamma$1

recent work introduces a second-order variational function $M_\gamma$2 built from proximal mappings and coderivatives. Under MSCQ, parabolic regularity, and additional verifiable conditions, tilt stability is characterized by point-based and neighborhood-based inequalities involving

$M_\gamma$3

yielding a no-gap second-order theory in which the sufficient and necessary conditions differ only by strict versus non-strict inequality (Mordukhovich et al., 15 Jul 2025).

For matrix optimization with smooth plus spectral structure,

$M_\gamma$4

tilt stability can be characterized through quadratic bundles. A minimal quadratic bundle exists for a broad class of spectral functions and is given explicitly by

$M_\gamma$5

This yields an SSOSC-type equivalence: $M_\gamma$6 is tilt-stable if and only if

$M_\gamma$7

for all nonzero $M_\gamma$8 in the affine critical cone (Ding et al., 5 Mar 2025).

5. Algorithmic consequences

Tilt stability is not only structural; it is also algorithmically useful. In nonsmooth optimization, two generalized Newton methods have been designed to converge specifically to tilt-stable local minimizers. For $M_\gamma$9 objectives, one algorithm uses the coderivative of the gradient mapping and the other uses the graphical derivative. Near a tilt-stable local minimizer, both methods have well-posed subproblems, and under semismooth* assumptions their iterates converge Q-superlinearly: $v=0$0 The same framework extends to continuously prox-regular functions by passing to the Moreau envelope, and to constrained optimization through extended-real-valued formulations and second subderivatives (Mordukhovich et al., 2020).

The role of tilt stability in these methods is explicit. It ensures stable inversion of generalized second-order objects, nonemptiness and compactness of Newton direction sets, and strong convexity or uniqueness of the local subproblems used to compute the step (Mordukhovich et al., 2020). In this sense, tilt-stable minimizers are the nonsmooth analog of nondegenerate solutions for classical Newton theory.

Growth characterizations also feed directly into first-order or proximal-type schemes. From the local $v=0$1-growth condition at a strict local minimizer, a proximal point algorithm of the form

$v=0$2

satisfies $v=0$3 and $v=0$4, with explicit geometric-type estimates for the distance and function-value errors (Corella et al., 2024). This places tilt-type stability within a broader convergence-rate theory driven by local growth.

6. Analogues, applications, and conceptual boundaries

A recurrent misconception is that stability requires global minimality. In the NLS on star graphs, the symmetric standing wave is orbitally stable for every admissible mass because it is a strict local minimizer of the constrained energy, even when it is not a ground state (Adami et al., 2015). For mass-critical NLS on non-compact metric graphs, constrained local minimizers likewise exist in regimes where the global infimum is not attained, and these local minimizers are identified as the variational objects relevant to orbital stability (Pierotti et al., 2019).

Another distinction concerns problems whose stability theory is analogous to tilt stability but not formulated in the Poliquin–Rockafellar sense. For the Mumford–Shah functional, a regular critical pair with positive definite second variation is an isolated local minimizer in the $v=0$5-topology (Bonacini et al., 2013). For the periodic Ohta–Kawasaki functional, strict positivity of second variation modulo translations yields isolated local minimality and a quadratic coercivity estimate with respect to a translation-invariant $v=0$6-distance (Cristoferi, 2015). These are strong local stability results, but they live in free-discontinuity and geometric variational settings rather than in the standard argminimum-mapping framework.

A further conceptual boundary separates local tilt stability from global uniqueness of tilted minimization problems. In a locally convex Hausdorff space, uniqueness of the global minimizer of

$v=0$7

for all tilts is characterized by essential strict convexity of the biconjugate $v=0$8 together with agreement of $v=0$9 and $M_\gamma(0)=\bar x$0 on $M_\gamma(0)=\bar x$1 (Ruf et al., 2021). This is closely related in spirit, but it addresses a global convex-envelope uniqueness problem rather than the local Lipschitz stability of a selected minimizer.

Taken together, these developments show that tilt-stable local minimizers form a central organizing notion in modern variational analysis. They unify perturbation stability, local growth, subdifferential regularity, and second-order curvature, and they supply a common language for problems ranging from weakly qualified nonlinear programming and nonpolyhedral conic optimization to nonsmooth Newton methods and local variational stability in PDE and geometric models (Drusvyatskiy et al., 2013, Mordukhovich et al., 2020, Liu et al., 2024).